Future work - quality, high resolution, bigger masks about edge-connect HOT 2 CLOSED

@fskurniak Thank you for your interest in our work.

1- Actually we are working on some new ideas that can increase the model capacity and performance! Adversarial Hinge loss, for example, works much better than the non-saturating loss for edge generator! And it's early to say but, multi-discriminator approach also seems to be working for us. We will soon publish our findings, so stay tuned!

2- The second stage of our model (the image completion model) has the capacity to inpaint images up-to 512x512 with a very good quality if it receives good edge information. As you can see the bottleneck of the model is the edge generator. To improve the quality of the edge-generator, we can play with the loss. But it's more important to increase the model receptive field. 512x512 images are very challenging and the model needs a good receptive field to cover the entire missing area! We are also developing some techniques to improve this problem. To answer your question, it always helps to continue training. In fact, when we were training the models, we always started with 128x128 and then use the wights for 256x256. Smaller images help early layers of the network learn faster.

3- As I mentioned earlier, bigger masks are very challenging. One approach to look into is an image pyramid. This paper by Nvidia is a good start.

from edge-connect.

1 - I will try experimenting with hinge loss, thanks. It might be interesting to see your setup especially in context of papers (deepfillv1 -> deepfillv2)
2 - thanks for the remarks about receptive field. Indeed this might be crucial
3 - interesting approach 👍

from edge-connect.